The growing complexity of data science leads to black box solutions that few people in an organization understand. You often hear about the difficulty of interpretability—explaining how an analytic model works—and that you need it to deploy models. But people use many black boxes without understanding them…if they’re reliable. It’s when the black box becomes unreliable that people lose trust. Mistrust is more likely to be created by the lack of reliability, and the lack of reliability is often the result of misunderstanding essential elements of analytics infrastructure and practice. The concept of reproducibility—the ability to get the same results given the same information—extends your view to include the environment and the data used to build and execute models. Mark Madsen examines reproducibility and the areas that underlie production analytics and explores the most frequently ignored and yet most essential capability, data management. The industry needs to consider its practices so that systems are more transparent and reliable, improving trust and increasing the likelihood that your analytic solutions will succeed. This talk will treat the black boxed of ML the way management perceives them, as black boxes. There is much work on explainable models, interpretability, etc. that are important to the task of reproducibility. Much of that is relevant to the practitioner, but the practitioner can become too focused on the part they are most familiar with and focused on. Reproducing the results needs more.

1. Copyright Third Nature, Inc.
Executive Briefing:
The black box –
interpretability,
reproducibility, and
data management
O’Reilly Artificial Intelligence
conference
London
October, 2019
Mark Madsen

2. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
What does management care about?
The perspective and problems
of the person responsible for
oversight of the results is
spread across the organization
and across multiple projects
▪ The owner of a decision or the
outcome of the decision. Aka the
process owner
▪ The CAO, CDO, VP of analytics,
aka “your boss” if you’re a data
scientist.

3. Copyright Third Nature, Inc.
Repeatability

4. Copyright Third Nature, Inc.
Operational predictability

5. Reproducibility
Can you support
the answer at a
later date?
Can you
replicate the
process used to
get the answer?
Can you get
similar results
on the same or
similar input?

6. Copyright Third Nature, Inc.
Scientific
reproducibility
Can you get the same
results given the same
starting conditions?
▪ Duplicate as much as
possible the experiment.
▪ Results of experimental
replication may not be the
same for many reasons
▪ Detailed statistical analysis
is needed
▪ And critical assessment of
experiments

7. Copyright Third Nature, Inc.
Reproducibility in
data science
Can you get the same
results on the same
data?
▪ Direct replication is
expected
▪ Assumption is that
same input = same
output
▪ You can have this with
an unexplainable box
▪ But there are also
confounding factors

8. Copyright Third Nature, Inc.
Moving from predictable rule-based systems to complex
mathematical systems, and from there to systems that
exhibit stochasticity, makes the task harder.
One thing worse than
a black box is a
random black box.

9. Copyright Third Nature, Inc.
If you trust the box then it’s ok not to understand it.
Without reproducibility there cannot be trust.
Trust comes from it working
It works = reliability, working over time
Reliability of your solution is a systems problem, not a
just a model problem (in the GST sense of system)
Therefore, if there is trust, or there is low risk, you
may not need to worry about reproducibility

10. Copyright Third Nature, Inc.
Interpretability and reproducibility are driven by trust,
which only matters when there is enough risk
Regulation and compliance
Material decisions
Big penalties
Complication: the cost of false
positives and false negatives
are usually different, so model
characteristics matter here.
Cost of error
Frequency of
decision
Low High
High
Don’t care
Oh crap

11. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Beyond toy examples, this problem matters
Edge case error problems will limit
AI applicability until they are solved
(don’t hold your breath).
The uncertainty challenge means:
▪ Calculate error costs and apply them
to your model before (and after)
▪ The more risk averse, the more
predictability is desire, the less
variance one can tolerate

12. Copyright Third Nature, Inc.
The real questions
Can I support this result at a later date?
▪ Do I care?
• Is the risk (cost) worth worrying about?
• Is the cost of reproducibility less than the risk and cost?
▪ Do I only need to justify the decision?
• interpretability and trust may be enough
• Unless there are audits or legal discovery involved
▪ Do I need to reproduce the results?
• May not need interpretability
• Need a lot of other things

13. Copyright Third Nature, Inc.
The real need is trust. Our trust is
based on all the elements that
are involved, not just the model.
The higher the stakes the more
you must think about all the ways
it could go wrong.
Reliability and robustness of the
technology environment is as
important as the model.
Reproducibility is not just the
data scientist’s problem – it is an
operational concern.

14. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Three categories of use, with differing trust levels
Embedded
Human in
the loop
Ad hoc

15. Copyright Third Nature, Inc.
Explainability only addresses the functioning of the model

16. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Three categories of use, with differing trust levels
Embedded
Human in
the loop
Ad hoc
In general, explainability is
more useful when there is
human agency and control in
decisions.
Reproducibility is increasingly
important where the machine
has agency because the
impact of automation expands
the scope and increases the
complexity beyond the model.

17. Machine learning is the smallest part of the environment
ML
Code
Analysis Tools
Data
Collection
Machine
Resource
Management
Serving
Infrastructure
Feature
Extraction
Configuration
Data
Verification
Process
Management
Tools
Monitoring
https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems
©2018 Teradata

18. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
The end result in the data science landscape is complexity
You can explain the model, but can you explain this?

19. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
ML Principle: CACE, Change Anything Change Everything
In embedded or
autonomous ML
everything is
connected.
Events usually
happen in real time.
ML is very sensitive
to context and input.
“So what if I
changed NumPy
in dev?”

20. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Working backwards from the decision or answer,
what do you need to have reproducibiity?
ProcessSource Run Env

21. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Fine for a single-run model. What if it’s
embedded in an application somewhere?
ProcessSource Run Env
App EnvProcess
Deploy

22. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
If a model is used over time, you will likely
change it. Or sources, or processes will change.
ProcessSource Run Env
App EnvProcess
Deploy
Run Env
Run Env
Process
Process
Run Env
Process Run Env
Process

23. Copyright Third Nature, Inc.
We’re so focused on the light switch that
we’re not talking about the light.

24. Copyright Third Nature, Inc.
Most emphasis in the
industry is on code and
code artifacts:
▪ Model repositories
▪ Model management
▪ Pipeline frameworks
▪ Packaging
▪ Versioning
▪ Tools
Why? Because vendors want
to sell you products for the
problem they helped create.
Managing the code
without managing
the data?

25. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
TIDY DATA: Hadley Wickham makes the case for
Tidy data sets, that have specific structure, are easy to
work with, that free analysts from mundane data
manipulation chores – there’s no need to start from
scratch and reinvent new methods for data cleaning
Source: Tidy Data by Hadley Wickham, Journal of Statistical Software, Vol 59, issue 10 (2014)
https://vita.had.co.nz/papers/tidy-data.html
What do the experts say?

26. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Andrew Ng
Leading AI Researcher
“If your boss asks you, tell
them that I said build a unified
data warehouse”
What about AI? GIGO!
Everyone wants shortcuts. There’s aren’t any shortcuts

27. Copyright Third Nature, Inc.
Reproducibility? No problem. Just save versions of data

28. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Today’s market solution:
the Data Lake*
*Depiction for illustrative
purposes only, no warranty
express or implied, actual
system may vary.
By a lot.

29. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
Today’s market solution:
the Data Lake + DIY
Data hoarding is not a data
management strategy

30. Copyright Third Nature, Inc.
A standard data science approach (to avoid)
One-Pipeline-
Per-Process
Redundant Effort /
Cost / Complexity /
etc.
WELL
HEAD
WELL
HEAD
Example use cases

31. Copyright Third Nature, Inc.
A dystopian model: Lake + data engineers + pipelines
The Lake with pipelines to files or tables for each model
is exactly the same pattern as mainframe COBOL batch
Data pipeline
code
We already know that people don’t scale. Don’t do this

32. Copyright Third Nature, Inc.
Self-service tradeoffs
Pay now or pay later,
but you will always pay.
The question is which
payment will be less.
Self-service gives
flexibility and agility, but
can reduce repeatability
and add duplicated
effort and conflicts, and
an increase in risk.

33. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
But that self-serve bit only applies to building models.
Embedding them in the enterprise is more involved
ProcessSource Run Env
App EnvProcess
Deploy
Run Env
Run Env
Process
Process
Run Env
Process Run Env
Process
Managing the data
on an ongoing
basis is a complex
problem to solve

34. Copyright Third Nature, Inc.
The organization-wide focus needs to be on
repeatability – where it can be supported

35. Copyright Third Nature, Inc.
Buildings above: flexibility, repurposing,
faster change above, funded independently
Applications
Utilities below: stability, reuse, slow
predictable change below, funded centrally
Infrastructure
The infrastructure is a hidden combination
of technology, process, and methods.
There is a need to draw boundaries for
responsibilities in the organization. It can’t
be all business or all IT.
Separate the uses from the infrastructure
– focus on the city, not the buildings

36. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
2855
Out of Deviation
Sensor Readings
Fraud
Events
CUSTOMER
EXPERIENCE
SALES
MARKETING
OPERATIONS
Abandoned
Carts
Online Price
Quotes
Social Media
Influencers
FINANCE
11234
Sensor data
formatting
Unit
Normalization
Vehicle/version
normalization
Make/model/year
selection
Data pipelines, data architecture, and curation

37. Copyright Third Nature, Inc.
Data curation is not data modeling
The problem with so many
sources, types, formats and
latencies of data is that it is now
impossible to create in advance
one model for all of it.
Data modeling is about the inside
of a dataset. Curation is about
the entire dataset.
It’s about: creating, labeling,
organizing, finding, navigating,
retiring.
Data curation, rather than data
modeling, is becoming the most
important data management
practice.

38. Copyright Third Nature, Inc.
Data can be maintained at multiple levels: not raw or DW
Ingredients
Goal: available
User needs a recipe
in order to make
use of the data.
Pre-mixed
Goal: discoverable
and integrateable
User needs a menu
to choose from the
data available
Meals
Goal: usable
User needs utensils
but is given a
finished meal

39. Copyright Third Nature, Inc.
Data architecture dictates technology and process
Collection
Capture metadata
(including request)
Record the structure
Apply keys
PII masking, restrict
Start lineage
Distribution
Common structures
RDM and MDM
Subject models
Make data findable
and linkable
Data provisioning
Consumption
Model for target use
Quality rules
Track provenance
Apply SLAs
As few engines as
possible – not fewer
Decide on policies for when to place data in which area

40. Copyright Third Nature, Inc.Copyright Third Nature, Inc.
ML has a lot of data requirements: no shortcuts
https://hackernoon.com/the-ai-hierarchy-of-needs-18f111fcc007

41. Copyright Third Nature, Inc.
Manage your data
(or it will manage you)
Data is the foundation of the
models, and the results. Focus on
managing data as well as the
code and tool artifacts.
Reproducibility, and data science
in general, requires that you
treat the operating and the build
environments as a whole.
Data science in the long term is
not about individual projects, but
organizational capability. This
means infrastructure, process,
governance.

42. Copyright Third Nature, Inc.
Things you can do
1. Determine what level of risk is acceptable based on the costs of
false positives and false negatives, or poor accuracy.
2. Manage the data history such that you can always reconstruct it.
3. Think through the dependencies outside the model that affect
reproducibility.
4. Track the configuration of everything in the chain of
dependencies.
5. Do not let a thousand flowers bloom with BYOT. You will have to
cut many of them due to complexity.
6. Do not let IT overcomplicate your environment with technology
because of belief things must be big. Brittle infrastructure can’t
be tolerated with embedded machine learning.
7. Make a rational decision about how much effort to expend.

43. Mark Madsen is a Fellow at Teradata in
the Technology and Innovation Office.
he focused on the data and analytics
ecosystem, problems of large-scale
design, and R&D.
Prior to that he was president of Third
Nature, a research and consulting firm
focused on analytics, integration and
data management.
Mark is an award-winning author,
architect and CTO whose work has
been featured in numerous industry
publications. He received awards for
his work from the American
Productivity & Quality Center and the
Smithsonian Institute. He is an
international speaker, chairs several
conferences and program committees.
Mark Madsen